LLM accuracy
Average of GSM8K, IFEval and MMLU-Pro. 56 Embedl builds of 26 models on 4 devices: which are the best trade-offs between speed and accuracy, and which you can run now.
Throughput against Average of 3
- Best trade-offs
- Other builds
Every build as a list
Try the best trade-offs on our device farm
Builds on this frontier with a device online that can take them now: the same run as on the model page.
The best trade-offs on every device
No other build here is both faster and more accurate than these. Fastest first; each row opens its build.
| # | Model | Device | Build | Throughput | Average of 3 | Speed-up |
|---|---|---|---|---|---|---|
| 1 | Qwen3.5 0.8B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 303.9 tok/s | 46.6% | 2.25× faster |
| 2 | Qwen3 0.6B | AGX Thor | Embedl-optimized · FlashHead · TensorRT Edge-LLM | 266.0 tok/s | 46.8% | 1.75× faster |
| 3 | Cosmos-Reason2 2B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 149.5 tok/s | 58.9% | 2.35× faster |
| 4 | Gemma 4 E2B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 120.0 tok/s | 74.9% | 2.57× faster |
| 5 | Gemma 4 E4B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 82.8 tok/s | 80.2% | 3.42× faster |
| 6 | Qwen3.6 35B-A3B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 82.1 tok/s | 85.0% | — |
| 7 | Gemma 4 26B-A4B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 58.4 tok/s | 89.3% | — |
| 8 | Gemma 4 31B | AGX Thor | Embedl-optimized · TensorRT Edge-LLM | 18.1 tok/s | 91.2% | — |